Face Generation¶

In this project, you'll use generative adversarial networks to generate new images of faces.

Get the Data¶

You'll be using two datasets in this project:

MNIST
CelebA

Since the celebA dataset is complex and you're doing GANs in a project for the first time, we want you to test your neural network on MNIST before CelebA. Running the GANs on MNIST will allow you to see how well your model trains sooner.

If you're using FloydHub, set data_dir to "/input" and use the FloydHub data ID "R5KrjnANiKVhLWAkpXhNBe".

data_dir = './data'

# FloydHub - Use with data ID "R5KrjnANiKVhLWAkpXhNBe"
#data_dir = '/input'


"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
import helper

helper.download_extract('mnist', data_dir)
helper.download_extract('celeba', data_dir)

Found mnist Data
Found celeba Data

Explore the Data¶

MNIST¶

As you're aware, the MNIST dataset contains images of handwritten digits. You can view the first number of examples by changing show_n_images.

show_n_images = 25

"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
%matplotlib inline
import os
from glob import glob
from matplotlib import pyplot

mnist_images = helper.get_batch(glob(os.path.join(data_dir, 'mnist/*.jpg'))[:show_n_images], 28, 28, 'L')
pyplot.imshow(helper.images_square_grid(mnist_images, 'L'), cmap='gray')

<matplotlib.image.AxesImage at 0x1bc9e135630>

CelebA¶

The CelebFaces Attributes Dataset (CelebA) dataset contains over 200,000 celebrity images with annotations. Since you're going to be generating faces, you won't need the annotations. You can view the first number of examples by changing show_n_images.

show_n_images = 25

"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
mnist_images = helper.get_batch(glob(os.path.join(data_dir, 'img_align_celeba/*.jpg'))[:show_n_images], 28, 28, 'RGB')
pyplot.imshow(helper.images_square_grid(mnist_images, 'RGB'))

<matplotlib.image.AxesImage at 0x1bc9e30cba8>

Preprocess the Data¶

Since the project's main focus is on building the GANs, we'll preprocess the data for you. The values of the MNIST and CelebA dataset will be in the range of -0.5 to 0.5 of 28x28 dimensional images. The CelebA images will be cropped to remove parts of the image that don't include a face, then resized down to 28x28.

The MNIST images are black and white images with a single color channel while the CelebA images have 3 color channels (RGB color channel).

Build the Neural Network¶

You'll build the components necessary to build a GANs by implementing the following functions below:

model_inputs
discriminator
generator
model_loss
model_opt
train

Check the Version of TensorFlow and Access to GPU¶

This will check to make sure you have the correct version of TensorFlow and access to a GPU

"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
from distutils.version import LooseVersion
import warnings
import tensorflow as tf

# Check TensorFlow Version
assert LooseVersion(tf.__version__) >= LooseVersion('1.0'), 'Please use TensorFlow version 1.0 or newer.  You are using {}'.format(tf.__version__)
print('TensorFlow Version: {}'.format(tf.__version__))

# Check for a GPU
if not tf.test.gpu_device_name():
    warnings.warn('No GPU found. Please use a GPU to train your neural network.')
else:
    print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))

TensorFlow Version: 1.0.1
Default GPU Device: /gpu:0

Input¶

Implement the model_inputs function to create TF Placeholders for the Neural Network. It should create the following placeholders:

Real input images placeholder with rank 4 using image_width, image_height, and image_channels.
Z input placeholder with rank 2 using z_dim.
Learning rate placeholder with rank 0.

Return the placeholders in the following the tuple (tensor of real input images, tensor of z data)

import problem_unittests as tests
import importlib
importlib.reload(tests)

def model_inputs(image_width, image_height, image_channels, z_dim):
    """
    Create the model inputs
    :param image_width: The input image width
    :param image_height: The input image height
    :param image_channels: The number of image channels
    :param z_dim: The dimension of Z
    :return: Tuple of (tensor of real input images, tensor of z data, learning rate)
    """
    # TODO: Implement Function
    inputs_real = tf.placeholder(tf.float32, [None, image_width, image_height, image_channels], name='input_real')
    inputs_z = tf.placeholder(tf.float32, [None, z_dim], name='input_z')
    learning_rate = tf.placeholder(tf.float32, name='learning_rate')

    return inputs_real, inputs_z, learning_rate


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_inputs(model_inputs)

Tests Passed

Discriminator¶

Implement discriminator to create a discriminator neural network that discriminates on images. This function should be able to reuse the variables in the neural network. Use tf.variable_scope with a scope name of "discriminator" to allow the variables to be reused. The function should return a tuple of (tensor output of the generator, tensor logits of the generator).

def lrelu(tensor, alpha=0.2):
    """
    Leaky ReLU implementation
    :param tensor: Input Tensor
    :param alpha: Leak scaling for negative inputs
    :return: Leaky ReLU output
    """
    return tf.maximum(tensor * alpha, tensor)

def discriminator(images, reuse=False, alpha=0.2, training=True):
    """
    Create the discriminator network
    :param image: Tensor of input image(s)
    :param reuse: Boolean if the weights should be reused
    Optional
    :param alpha: Leak factor for Leaky ReLU
    :param training: Boolean for batch normalization to use batch or population statistics
    :return: Tuple of (tensor output of the discriminator, tensor logits of the discriminator)
    """
    # TODO: Implement Function
    # input is 28x28x1 or 28x28x3
    with tf.variable_scope('discriminator', reuse=reuse):
        x = tf.layers.conv2d(images, filters=64, kernel_size=5, strides=2, padding='same', activation=None)
        x = lrelu(x, alpha)
        # Now 14x14x64
        
        # 2 conv layers for 7x7x128 -> 4x4x256
        for filter_depth in [128, 256]:
            x = tf.layers.conv2d(x, filters=filter_depth, kernel_size=5, strides=2, padding='same', activation=None, use_bias=False)
            # Use batch normalization
            x = tf.layers.batch_normalization(x, training=training)
            x = lrelu(x, alpha)
        # Now 4x4x256
        
        flat = tf.reshape(x, [-1, 4*4*256])
        logits = tf.layers.dense(flat, 1, activation=None)
        out = tf.sigmoid(logits)
        #print(x.shape, flat.shape, logits.shape)

    return out, logits


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_discriminator(discriminator, tf)

Tests Passed

Generator¶

Implement generator to generate an image using z. This function should be able to reuse the variables in the neural network. Use tf.variable_scope with a scope name of "generator" to allow the variables to be reused. The function should return the generated 28 x 28 x out_channel_dim images.

def generator(z, out_channel_dim, is_train=True, alpha=0.2):
    """
    Create the generator network
    :param z: Input z
    :param out_channel_dim: The number of channels in the output image
    :param is_train: Boolean if generator is being used for training
    Optional
    :param alpha: Leak factor for Leaky ReLU
    :return: The tensor output of the generator
    """
    # TODO: Implement Function
    # during training, don't reuse, during generation, reuse variables
    with tf.variable_scope('generator', reuse=not is_train):
        # Fully connected layer
        x = tf.layers.dense(z, 2*2*1024, activation=None, use_bias=False)
        # Reshape to start the conv stack
        x = tf.reshape(x, (-1, 2, 2, 1024))
        x = lrelu(x, alpha)
        # Now 2x2x1024
        
        x = tf.layers.conv2d_transpose(x, 512, kernel_size=5, strides=2, padding='valid', activation=None, use_bias=False)
        x = tf.layers.batch_normalization(x, training=is_train)
        x = lrelu(x, alpha)
        # 7x7x512
        
        #print(x.shape)
        x = tf.layers.conv2d_transpose(x, 256, kernel_size=5, strides=2, padding='same', activation=None, use_bias=False)
        x = tf.layers.batch_normalization(x, training=is_train)
        x = lrelu(x, alpha)
        # 14x14x256
        
        #print(x.shape)
        logits = tf.layers.conv2d_transpose(x, out_channel_dim, kernel_size=5, strides=2, padding='same')
        # 28x28xout_channel_dim
        
        #print(logits.shape)
        # the real images from get_batches are scaled between -0.5 and 0.5
        #scale = 0.5
        #out = tf.multiply(tf.tanh(logits), scale)
        out = tf.tanh(logits)
    
    return out


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_generator(generator, tf)

Tests Passed

Loss¶

Implement model_loss to build the GANs for training and calculate the loss. The function should return a tuple of (discriminator loss, generator loss). Use the following functions you implemented:

discriminator(images, reuse=False)
generator(z, out_channel_dim, is_train=True)

def model_loss(input_real, input_z, out_channel_dim, alpha=0.2, label_smooth=0.1):
    """
    Get the loss for the discriminator and generator
    :param input_real: Images from the real dataset
    :param input_z: Z input
    :param out_channel_dim: The number of channels in the output image
    Optional
    :param alpha: Leak factor for Leaky ReLU
    :param label_smooth: Label smoothing for better discriminator generalization
    :return: A tuple of (discriminator loss, generator loss)
    """
    # TODO: Implement Function
    g_model = generator(input_z, out_channel_dim, alpha=alpha)
    d_model_real, d_logits_real = discriminator(input_real, alpha=alpha)
    d_model_fake, d_logits_fake = discriminator(g_model, reuse=True, alpha=alpha)
    
    d_loss_real = tf.reduce_mean(
        tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real, labels=tf.ones_like(d_model_real) * (1 - label_smooth)))
    d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.zeros_like(d_model_fake)))
    g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.ones_like(d_model_fake)))
    
    d_loss = d_loss_real + d_loss_fake
    
    return d_loss, g_loss


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_loss(model_loss)

Tests Passed

Optimization¶

Implement model_opt to create the optimization operations for the GANs. Use tf.trainable_variables to get all the trainable variables. Filter the variables with names that are in the discriminator and generator scope names. The function should return a tuple of (discriminator training operation, generator training operation).

def model_opt(d_loss, g_loss, learning_rate, beta1):
    """
    Get optimization operations
    :param d_loss: Discriminator loss Tensor
    :param g_loss: Generator loss Tensor
    :param learning_rate: Learning Rate Placeholder
    :param beta1: The exponential decay rate for the 1st moment in the optimizer
    :return: A tuple of (discriminator training operation, generator training operation)
    """
    # TODO: Implement Function
    # Get weights and bias
    t_vars = tf.trainable_variables()
    d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
    g_vars = [var for var in t_vars if var.name.startswith('generator')]
    
    # Get batch normalization variables to update
    d_update_opts = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='discriminator')
    g_update_opts = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='generator')
    
    with tf.control_dependencies(d_update_opts):
        d_train_opt = tf.train.AdamOptimizer(learning_rate, beta1=beta1).minimize(d_loss, var_list=d_vars)
    with tf.control_dependencies(g_update_opts):
        g_train_opt = tf.train.AdamOptimizer(learning_rate, beta1=beta1).minimize(g_loss, var_list=g_vars)
    
    
    return d_train_opt, g_train_opt


"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
tests.test_model_opt(model_opt, tf)

Tests Passed

Neural Network Training¶

Show Output¶

Use this function to show the current output of the generator during training. It will help you determine how well the GANs is training.

"""
DON'T MODIFY ANYTHING IN THIS CELL
"""
import numpy as np

def show_generator_output(sess, n_images, input_z, out_channel_dim, image_mode):
    """
    Show example output for the generator
    :param sess: TensorFlow session
    :param n_images: Number of Images to display
    :param input_z: Input Z Tensor
    :param out_channel_dim: The number of channels in the output image
    :param image_mode: The mode to use for images ("RGB" or "L")
    """
    cmap = None if image_mode == 'RGB' else 'gray'
    z_dim = input_z.get_shape().as_list()[-1]
    example_z = np.random.uniform(-1, 1, size=[n_images, z_dim])

    samples = sess.run(
        generator(input_z, out_channel_dim, False),
        feed_dict={input_z: example_z})

    images_grid = helper.images_square_grid(samples, image_mode)
    pyplot.imshow(images_grid, cmap=cmap)
    pyplot.show()

Train¶

Implement train to build and train the GANs. Use the following functions you implemented:

model_inputs(image_width, image_height, image_channels, z_dim)
model_loss(input_real, input_z, out_channel_dim)
model_opt(d_loss, g_loss, learning_rate, beta1)

Use the show_generator_output to show generator output while you train. Running show_generator_output for every batch will drastically increase training time and increase the size of the notebook. It's recommended to print the generator output every 100 batches.

try:
    from tqdm import tqdm
except ImportError:
    def tqdm(x):
        return x

!mkdir checkpoints

A subdirectory or file checkpoints already exists.

def plot_losses(losses):
    """
    Plot training losses
    :param losses: list of tuples of (discriminator loss, generator loss)
    """
    import matplotlib.pyplot as plt
    fig, ax = plt.subplots()
    # each row in losses is (discriminator loss, generator loss)
    losses = np.array(losses)
    # transpose of losses would give all discriminator losses as first row and generator losses as second row
    plt.plot(losses.T[0], label='Discriminator', alpha=0.3)
    plt.plot(losses.T[1], label='Generator')
    plt.title('Training Losses')
    plt.legend()

def train(epoch_count, batch_size, z_dim, learning_rate, beta1, get_batches, data_shape, data_image_mode, alpha=0.2, 
          print_every=10, show_every=100, show_n_images=25):
    """
    Train the GAN
    :param epoch_count: Number of epochs
    :param batch_size: Batch Size
    :param z_dim: Z dimension
    :param learning_rate: Learning Rate
    :param beta1: The exponential decay rate for the 1st moment in the optimizer
    :param get_batches: Function to get batches
    :param data_shape: Shape of the data
    :param data_image_mode: The image mode to use for images ("RGB" or "L")
    Optional
    :param alpha: Leak factor for Leaky ReLU
    :param print_every: Print loss every n steps
    :param show_every: Show images generated every n steps
    :param show_n_images: Number of images to show each time
    """
    # TODO: Build Model
    
    # data_shape is (samples x width x height x channels)
    input_real, input_z, learn_rate = model_inputs(*data_shape[1:], z_dim)
    out_channel_dim = data_shape[-1]
    d_loss, g_loss = model_loss(input_real, input_z, out_channel_dim, alpha=alpha)
    d_opt, g_opt = model_opt(d_loss, g_loss, learn_rate, beta1)
    
    saver = tf.train.Saver()
    steps = 0
    losses = []
    
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        epoch_pbar = tqdm(range(epoch_count), desc='Epochs', unit='epoch')
        
        #for epoch_i in range(epoch_count):
        for epoch_i in epoch_pbar:
            batch_count = data_shape[0]//batch_size
            batch_pbar = tqdm(get_batches(batch_size), desc='Batches', unit='batch', miniters=show_every, 
                              total=batch_count, maxinterval=360)
            batch_i = 0
            #for batch_images in get_batches(batch_size):
            for batch_images in batch_pbar:
                # TODO: Train Model
                steps += 1
                batch_i += 1
                # the real images from get_batches are scaled between -0.5 and 0.5, so we scale to -1 to 1 to match generator
                batch_images *= 2
                # Sample random noise for Generator
                batch_z = np.random.uniform(-1, 1, size=(batch_size, z_dim))
                
                # Run optimizers
                _ = sess.run(d_opt, feed_dict={input_real: batch_images, input_z: batch_z, learn_rate: learning_rate})
                # Train generator twice to meet loss lower than discriminator requirement
                _ = sess.run(g_opt, feed_dict={input_z: batch_z, learn_rate: learning_rate})
                _ = sess.run(g_opt, feed_dict={input_z: batch_z, learn_rate: learning_rate})
                
                if steps % print_every == 0:
                    train_loss_d = d_loss.eval({input_real: batch_images, input_z: batch_z})
                    train_loss_g = g_loss.eval({input_z: batch_z})
                    
                    print('Epoch {}/{}...'.format(epoch_i, epoch_count),
                          'Batch {}/{}...'.format(batch_i, batch_count),
                          'Discriminator loss: {:.4f}'.format(train_loss_d),
                          'Generator loss: {:.4f}'.format(train_loss_g))
                    losses.append((train_loss_d, train_loss_g))
                
                if steps % show_every == 0:
                    show_generator_output(sess, show_n_images, input_z, out_channel_dim, data_image_mode)
        saver.save(sess, './checkpoints/generator_bs{}_zd{}_lr{}_b{}.ckpt'.format(batch_size, z_dim, learning_rate, beta1))
    plot_losses(losses)

MNIST¶

Test your GANs architecture on MNIST. After 2 epochs, the GANs should be able to generate images that look like handwritten digits. Make sure the loss of the generator is lower than the loss of the discriminator or close to 0.

mnist_dataset = helper.Dataset('mnist', glob(os.path.join(data_dir, 'mnist/*.jpg')))
mnist_dataset.shape

(60000, 28, 28, 1)

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
celeba_dataset.shape

(202599, 28, 28, 3)

batch_size = 128
z_dim = 100
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 2

mnist_dataset = helper.Dataset('mnist', glob(os.path.join(data_dir, 'mnist/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, mnist_dataset.get_batches,
          mnist_dataset.shape, mnist_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/2 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                    | 0/468 [00:00<?, ?batch/s]

Epoch 0/2... Batch 10/468... Discriminator loss: 0.4917 Generator loss: 5.2180
Epoch 0/2... Batch 20/468... Discriminator loss: 3.6643 Generator loss: 0.0700
Epoch 0/2... Batch 30/468... Discriminator loss: 2.1145 Generator loss: 3.2754
Epoch 0/2... Batch 40/468... Discriminator loss: 1.9406 Generator loss: 1.6231
Epoch 0/2... Batch 50/468... Discriminator loss: 0.8335 Generator loss: 1.9555
Epoch 0/2... Batch 60/468... Discriminator loss: 1.3331 Generator loss: 0.9221
Epoch 0/2... Batch 70/468... Discriminator loss: 1.4760 Generator loss: 0.5973
Epoch 0/2... Batch 80/468... Discriminator loss: 0.9745 Generator loss: 2.3600
Epoch 0/2... Batch 90/468... Discriminator loss: 0.8449 Generator loss: 1.4911
Epoch 0/2... Batch 100/468... Discriminator loss: 0.8251 Generator loss: 1.5158

Batches:  21%|███████████████▊                                                          | 100/468 [00:26<01:38,  3.73batch/s]

Epoch 0/2... Batch 110/468... Discriminator loss: 1.2290 Generator loss: 1.7586
Epoch 0/2... Batch 120/468... Discriminator loss: 1.5158 Generator loss: 0.6855
Epoch 0/2... Batch 130/468... Discriminator loss: 1.6074 Generator loss: 0.9875
Epoch 0/2... Batch 140/468... Discriminator loss: 2.0240 Generator loss: 0.3558
Epoch 0/2... Batch 150/468... Discriminator loss: 0.9545 Generator loss: 1.3825
Epoch 0/2... Batch 160/468... Discriminator loss: 1.1279 Generator loss: 1.2117
Epoch 0/2... Batch 170/468... Discriminator loss: 1.1458 Generator loss: 1.0389
Epoch 0/2... Batch 180/468... Discriminator loss: 1.1388 Generator loss: 1.1487
Epoch 0/2... Batch 190/468... Discriminator loss: 1.2500 Generator loss: 1.3857
Epoch 0/2... Batch 200/468... Discriminator loss: 1.7996 Generator loss: 0.7772

Batches:  43%|███████████████████████████████▌                                          | 200/468 [00:52<01:10,  3.80batch/s]

Epoch 0/2... Batch 210/468... Discriminator loss: 1.2502 Generator loss: 0.8111
Epoch 0/2... Batch 220/468... Discriminator loss: 1.1672 Generator loss: 1.0612
Epoch 0/2... Batch 230/468... Discriminator loss: 1.4713 Generator loss: 0.8002
Epoch 0/2... Batch 240/468... Discriminator loss: 1.1337 Generator loss: 1.2037
Epoch 0/2... Batch 250/468... Discriminator loss: 1.2027 Generator loss: 1.0628
Epoch 0/2... Batch 260/468... Discriminator loss: 1.2021 Generator loss: 1.3674
Epoch 0/2... Batch 270/468... Discriminator loss: 1.4274 Generator loss: 0.6277
Epoch 0/2... Batch 280/468... Discriminator loss: 1.2764 Generator loss: 1.0985
Epoch 0/2... Batch 290/468... Discriminator loss: 1.1557 Generator loss: 1.0927
Epoch 0/2... Batch 300/468... Discriminator loss: 1.2954 Generator loss: 0.8428

Batches:  64%|███████████████████████████████████████████████▍                          | 300/468 [01:19<00:44,  3.75batch/s]

Epoch 0/2... Batch 310/468... Discriminator loss: 1.3570 Generator loss: 1.0544
Epoch 0/2... Batch 320/468... Discriminator loss: 1.3732 Generator loss: 1.0085
Epoch 0/2... Batch 330/468... Discriminator loss: 1.3976 Generator loss: 1.1268
Epoch 0/2... Batch 340/468... Discriminator loss: 1.3836 Generator loss: 1.1496
Epoch 0/2... Batch 350/468... Discriminator loss: 1.2577 Generator loss: 1.3831
Epoch 0/2... Batch 360/468... Discriminator loss: 1.4038 Generator loss: 0.8038
Epoch 0/2... Batch 370/468... Discriminator loss: 1.4000 Generator loss: 1.0657
Epoch 0/2... Batch 380/468... Discriminator loss: 1.2514 Generator loss: 0.8665
Epoch 0/2... Batch 390/468... Discriminator loss: 1.4491 Generator loss: 0.6437
Epoch 0/2... Batch 400/468... Discriminator loss: 1.2428 Generator loss: 0.9599

Batches:  85%|███████████████████████████████████████████████████████████████▏          | 400/468 [01:44<00:17,  3.82batch/s]

Epoch 0/2... Batch 410/468... Discriminator loss: 1.4123 Generator loss: 1.3603
Epoch 0/2... Batch 420/468... Discriminator loss: 1.4848 Generator loss: 0.4702
Epoch 0/2... Batch 430/468... Discriminator loss: 1.3709 Generator loss: 0.9749
Epoch 0/2... Batch 440/468... Discriminator loss: 1.3116 Generator loss: 1.0834
Epoch 0/2... Batch 450/468... Discriminator loss: 1.2400 Generator loss: 1.0020
Epoch 0/2... Batch 460/468... Discriminator loss: 1.2240 Generator loss: 1.1380

batch_size = 128
z_dim = 200
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 2

mnist_dataset = helper.Dataset('mnist', glob(os.path.join(data_dir, 'mnist/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, mnist_dataset.get_batches,
          mnist_dataset.shape, mnist_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/2 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                    | 0/468 [00:00<?, ?batch/s]

Epoch 0/2... Batch 10/468... Discriminator loss: 0.4435 Generator loss: 10.7345
Epoch 0/2... Batch 20/468... Discriminator loss: 0.6819 Generator loss: 6.3631
Epoch 0/2... Batch 30/468... Discriminator loss: 2.0784 Generator loss: 0.2784
Epoch 0/2... Batch 40/468... Discriminator loss: 1.5113 Generator loss: 0.5470
Epoch 0/2... Batch 50/468... Discriminator loss: 1.9211 Generator loss: 2.1227
Epoch 0/2... Batch 60/468... Discriminator loss: 1.6089 Generator loss: 1.6311
Epoch 0/2... Batch 70/468... Discriminator loss: 1.1816 Generator loss: 2.3494
Epoch 0/2... Batch 80/468... Discriminator loss: 0.7240 Generator loss: 2.5696
Epoch 0/2... Batch 90/468... Discriminator loss: 1.6865 Generator loss: 2.2101
Epoch 0/2... Batch 100/468... Discriminator loss: 1.0282 Generator loss: 2.3086

Batches:  21%|███████████████▊                                                          | 100/468 [00:50<03:05,  1.98batch/s]

Epoch 0/2... Batch 110/468... Discriminator loss: 1.0414 Generator loss: 2.3266
Epoch 0/2... Batch 120/468... Discriminator loss: 0.8370 Generator loss: 1.2102
Epoch 0/2... Batch 130/468... Discriminator loss: 0.8832 Generator loss: 1.4083
Epoch 0/2... Batch 140/468... Discriminator loss: 1.0062 Generator loss: 3.5232
Epoch 0/2... Batch 150/468... Discriminator loss: 1.4988 Generator loss: 0.5457
Epoch 0/2... Batch 160/468... Discriminator loss: 1.2900 Generator loss: 1.2644
Epoch 0/2... Batch 170/468... Discriminator loss: 0.8047 Generator loss: 1.4298
Epoch 0/2... Batch 180/468... Discriminator loss: 1.2491 Generator loss: 0.8524
Epoch 0/2... Batch 190/468... Discriminator loss: 1.5369 Generator loss: 3.5946
Epoch 0/2... Batch 200/468... Discriminator loss: 1.2011 Generator loss: 1.0701

Batches:  43%|███████████████████████████████▌                                          | 200/468 [01:36<02:11,  2.03batch/s]

Epoch 0/2... Batch 210/468... Discriminator loss: 1.1852 Generator loss: 0.9097
Epoch 0/2... Batch 220/468... Discriminator loss: 1.2709 Generator loss: 1.9000
Epoch 0/2... Batch 230/468... Discriminator loss: 1.5873 Generator loss: 0.6840
Epoch 0/2... Batch 240/468... Discriminator loss: 0.9993 Generator loss: 1.4596
Epoch 0/2... Batch 250/468... Discriminator loss: 1.2687 Generator loss: 1.4305
Epoch 0/2... Batch 260/468... Discriminator loss: 1.1471 Generator loss: 1.2122
Epoch 0/2... Batch 270/468... Discriminator loss: 1.0257 Generator loss: 1.2598
Epoch 0/2... Batch 280/468... Discriminator loss: 1.0547 Generator loss: 1.1313
Epoch 0/2... Batch 290/468... Discriminator loss: 1.3830 Generator loss: 0.6868
Epoch 0/2... Batch 300/468... Discriminator loss: 0.9425 Generator loss: 1.2102

Batches:  64%|███████████████████████████████████████████████▍                          | 300/468 [02:24<01:21,  2.05batch/s]

Epoch 0/2... Batch 310/468... Discriminator loss: 1.3713 Generator loss: 1.1332
Epoch 0/2... Batch 320/468... Discriminator loss: 1.1907 Generator loss: 1.0091
Epoch 0/2... Batch 330/468... Discriminator loss: 1.4063 Generator loss: 1.2298
Epoch 0/2... Batch 340/468... Discriminator loss: 1.6128 Generator loss: 1.0067
Epoch 0/2... Batch 350/468... Discriminator loss: 1.5502 Generator loss: 0.8953
Epoch 0/2... Batch 360/468... Discriminator loss: 1.4479 Generator loss: 1.0112
Epoch 0/2... Batch 370/468... Discriminator loss: 1.1682 Generator loss: 1.5175
Epoch 0/2... Batch 380/468... Discriminator loss: 1.3367 Generator loss: 0.8333
Epoch 0/2... Batch 390/468... Discriminator loss: 1.3867 Generator loss: 1.1652
Epoch 0/2... Batch 400/468... Discriminator loss: 1.3632 Generator loss: 1.0347

Batches:  85%|███████████████████████████████████████████████████████████████▏          | 400/468 [03:13<00:33,  2.05batch/s]

Epoch 0/2... Batch 410/468... Discriminator loss: 1.3535 Generator loss: 1.2404
Epoch 0/2... Batch 420/468... Discriminator loss: 1.7433 Generator loss: 0.3725
Epoch 0/2... Batch 430/468... Discriminator loss: 1.4140 Generator loss: 1.2692
Epoch 0/2... Batch 440/468... Discriminator loss: 1.3175 Generator loss: 0.8893
Epoch 0/2... Batch 450/468... Discriminator loss: 1.3028 Generator loss: 0.9682
Epoch 0/2... Batch 460/468... Discriminator loss: 1.2309 Generator loss: 1.0571

CelebA¶

Run your GANs on CelebA. It will take around 20 minutes on the average GPU to run one epoch. You can run the whole epoch or stop when it starts to generate realistic faces.

batch_size = 128
z_dim = 100
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2
# Generator Losses seem ok, but faces generated are bad

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

batch_size = 64
z_dim = 100
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2
# Generator Losses seem ok, but faces generated are bad

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

batch_size = 32
z_dim = 100
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2
# Generator Losses seem ok, but faces generated are bad

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

batch_size = 32
z_dim = 200 # Try doubling from earlier attempt of 100
learning_rate = 0.0005
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

batch_size = 256
z_dim = 200
learning_rate = 0.01 # too high
beta1 = 0.5
alpha = 0.2
# Generator Losses jumping around too much. Batch size and learning rate seem too high to be stable. Faces are too noisy

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                    | 0/791 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/791... Discriminator loss: 3.4570 Generator loss: 0.1437
Epoch 0/1... Batch 20/791... Discriminator loss: 2.4284 Generator loss: 3.3371
Epoch 0/1... Batch 30/791... Discriminator loss: 0.6787 Generator loss: 5.0026
Epoch 0/1... Batch 40/791... Discriminator loss: 1.2642 Generator loss: 6.0321
Epoch 0/1... Batch 50/791... Discriminator loss: 2.6873 Generator loss: 1.6424
Epoch 0/1... Batch 60/791... Discriminator loss: 0.6704 Generator loss: 2.7082
Epoch 0/1... Batch 70/791... Discriminator loss: 1.0521 Generator loss: 2.3508
Epoch 0/1... Batch 80/791... Discriminator loss: 1.3881 Generator loss: 1.0394
Epoch 0/1... Batch 90/791... Discriminator loss: 0.6054 Generator loss: 2.5848
Epoch 0/1... Batch 100/791... Discriminator loss: 0.5554 Generator loss: 1.8850

Batches:  13%|█████████▎                                                                | 100/791 [02:05<14:24,  1.25s/batch]

Epoch 0/1... Batch 110/791... Discriminator loss: 0.4814 Generator loss: 3.1759
Epoch 0/1... Batch 120/791... Discriminator loss: 0.3945 Generator loss: 4.0385
Epoch 0/1... Batch 130/791... Discriminator loss: 0.9977 Generator loss: 1.0179
Epoch 0/1... Batch 140/791... Discriminator loss: 1.0327 Generator loss: 1.8289
Epoch 0/1... Batch 150/791... Discriminator loss: 1.0169 Generator loss: 1.8406
Epoch 0/1... Batch 160/791... Discriminator loss: 0.6310 Generator loss: 2.2822
Epoch 0/1... Batch 170/791... Discriminator loss: 1.0231 Generator loss: 1.4359
Epoch 0/1... Batch 180/791... Discriminator loss: 0.7165 Generator loss: 1.8134
Epoch 0/1... Batch 190/791... Discriminator loss: 0.4087 Generator loss: 3.2192
Epoch 0/1... Batch 200/791... Discriminator loss: 0.3802 Generator loss: 3.9742

Batches:  25%|██████████████████▋                                                       | 200/791 [04:05<12:10,  1.24s/batch]

Epoch 0/1... Batch 210/791... Discriminator loss: 0.3910 Generator loss: 3.1805
Epoch 0/1... Batch 220/791... Discriminator loss: 0.3506 Generator loss: 4.3192
Epoch 0/1... Batch 230/791... Discriminator loss: 0.4349 Generator loss: 4.6078
Epoch 0/1... Batch 240/791... Discriminator loss: 0.3470 Generator loss: 4.2532
Epoch 0/1... Batch 250/791... Discriminator loss: 0.3636 Generator loss: 3.7173
Epoch 0/1... Batch 260/791... Discriminator loss: 0.3403 Generator loss: 4.6121
Epoch 0/1... Batch 270/791... Discriminator loss: 0.3348 Generator loss: 5.3991
Epoch 0/1... Batch 280/791... Discriminator loss: 0.3412 Generator loss: 5.8729
Epoch 0/1... Batch 290/791... Discriminator loss: 0.3323 Generator loss: 6.4463
Epoch 0/1... Batch 300/791... Discriminator loss: 0.3302 Generator loss: 6.1946

Batches:  38%|████████████████████████████                                              | 300/791 [06:10<10:09,  1.24s/batch]

Epoch 0/1... Batch 310/791... Discriminator loss: 0.3865 Generator loss: 9.5156
Epoch 0/1... Batch 320/791... Discriminator loss: 0.3643 Generator loss: 15.2904
Epoch 0/1... Batch 330/791... Discriminator loss: 0.3617 Generator loss: 4.7436
Epoch 0/1... Batch 340/791... Discriminator loss: 0.3428 Generator loss: 5.5071
Epoch 0/1... Batch 350/791... Discriminator loss: 0.3278 Generator loss: 10.2527
Epoch 0/1... Batch 360/791... Discriminator loss: 0.3459 Generator loss: 4.5914
Epoch 0/1... Batch 370/791... Discriminator loss: 0.3911 Generator loss: 4.5637
Epoch 0/1... Batch 380/791... Discriminator loss: 0.3335 Generator loss: 5.2410
Epoch 0/1... Batch 390/791... Discriminator loss: 0.4299 Generator loss: 8.5181
Epoch 0/1... Batch 400/791... Discriminator loss: 2.4820 Generator loss: 4.2033

Batches:  51%|█████████████████████████████████████▍                                    | 400/791 [07:20<07:01,  1.08s/batch]

Epoch 0/1... Batch 410/791... Discriminator loss: 1.9930 Generator loss: 0.9494
Epoch 0/1... Batch 420/791... Discriminator loss: 0.9337 Generator loss: 1.5956
Epoch 0/1... Batch 430/791... Discriminator loss: 1.4076 Generator loss: 1.1299
Epoch 0/1... Batch 440/791... Discriminator loss: 1.0554 Generator loss: 1.0217
Epoch 0/1... Batch 450/791... Discriminator loss: 1.2807 Generator loss: 1.0521
Epoch 0/1... Batch 460/791... Discriminator loss: 1.3113 Generator loss: 0.9976
Epoch 0/1... Batch 470/791... Discriminator loss: 1.7538 Generator loss: 1.4148
Epoch 0/1... Batch 480/791... Discriminator loss: 1.3521 Generator loss: 0.9649
Epoch 0/1... Batch 490/791... Discriminator loss: 1.2591 Generator loss: 0.6668
Epoch 0/1... Batch 500/791... Discriminator loss: 1.1729 Generator loss: 1.0440

batch_size = 32
z_dim = 200
learning_rate = 0.001
beta1 = 0.5
alpha = 0.2
# the visuals on batch 6300 randomly looks acceptable, but unable to get similar results on second run
"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 0.5115 Generator loss: 3.7177
Epoch 0/1... Batch 20/6331... Discriminator loss: 1.4912 Generator loss: 1.4111
Epoch 0/1... Batch 30/6331... Discriminator loss: 1.7850 Generator loss: 0.8752
Epoch 0/1... Batch 40/6331... Discriminator loss: 1.9232 Generator loss: 0.5031
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.2939 Generator loss: 2.2244
Epoch 0/1... Batch 60/6331... Discriminator loss: 1.2366 Generator loss: 1.0862
Epoch 0/1... Batch 70/6331... Discriminator loss: 2.1008 Generator loss: 0.2211
Epoch 0/1... Batch 80/6331... Discriminator loss: 1.8862 Generator loss: 0.7034
Epoch 0/1... Batch 90/6331... Discriminator loss: 2.2137 Generator loss: 0.4439
Epoch 0/1... Batch 100/6331... Discriminator loss: 1.7722 Generator loss: 0.9343

Batches:   2%|█▏                                                                       | 100/6331 [00:14<14:59,  6.92batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 1.5820 Generator loss: 1.2259
Epoch 0/1... Batch 120/6331... Discriminator loss: 1.6295 Generator loss: 0.9286
Epoch 0/1... Batch 130/6331... Discriminator loss: 1.0056 Generator loss: 1.2434
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.9914 Generator loss: 0.3788
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.9521 Generator loss: 0.5895
Epoch 0/1... Batch 160/6331... Discriminator loss: 1.8908 Generator loss: 0.4599
Epoch 0/1... Batch 170/6331... Discriminator loss: 1.7976 Generator loss: 0.5415
Epoch 0/1... Batch 180/6331... Discriminator loss: 1.7391 Generator loss: 0.4484
Epoch 0/1... Batch 190/6331... Discriminator loss: 1.6702 Generator loss: 0.5741
Epoch 0/1... Batch 200/6331... Discriminator loss: 1.8568 Generator loss: 0.4284

Batches:   3%|██▎                                                                      | 200/6331 [00:28<14:35,  7.00batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.5538 Generator loss: 0.5793
Epoch 0/1... Batch 220/6331... Discriminator loss: 1.8365 Generator loss: 0.4889
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.6418 Generator loss: 0.6146
Epoch 0/1... Batch 240/6331... Discriminator loss: 2.0320 Generator loss: 0.3784
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.4182 Generator loss: 0.9006
Epoch 0/1... Batch 260/6331... Discriminator loss: 1.5828 Generator loss: 0.6806
Epoch 0/1... Batch 270/6331... Discriminator loss: 1.6587 Generator loss: 0.5697
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.3113 Generator loss: 0.8651
Epoch 0/1... Batch 290/6331... Discriminator loss: 1.5334 Generator loss: 0.6386
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.4307 Generator loss: 0.7520

Batches:   5%|███▍                                                                     | 300/6331 [00:41<14:00,  7.17batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.8979 Generator loss: 0.5056
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.5771 Generator loss: 0.7117
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.5796 Generator loss: 0.7230
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.6321 Generator loss: 0.6169
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.5606 Generator loss: 0.6677
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.7702 Generator loss: 0.6650
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.5790 Generator loss: 0.7038
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.4638 Generator loss: 0.7664
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.5293 Generator loss: 0.7090
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.5801 Generator loss: 0.6138

Batches:   6%|████▌                                                                    | 400/6331 [00:54<13:30,  7.31batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.7609 Generator loss: 0.6207
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.6922 Generator loss: 0.6316
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.6942 Generator loss: 0.6461
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.6414 Generator loss: 0.6087
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.6476 Generator loss: 0.6879
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.4302 Generator loss: 0.6643
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.5180 Generator loss: 0.7492
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.6794 Generator loss: 0.5900
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.4455 Generator loss: 0.7339
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.4423 Generator loss: 0.7312

batch_size = 32
z_dim = 200
learning_rate = 0.001
beta1 = 0.5
alpha = 0.2
# unable to match the visuals of the previous run for some reason

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 0.5044 Generator loss: 4.1206
Epoch 0/1... Batch 20/6331... Discriminator loss: 2.5716 Generator loss: 1.7344
Epoch 0/1... Batch 30/6331... Discriminator loss: 1.3623 Generator loss: 2.3181
Epoch 0/1... Batch 40/6331... Discriminator loss: 1.1700 Generator loss: 2.5861
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.2700 Generator loss: 2.1419
Epoch 0/1... Batch 60/6331... Discriminator loss: 1.5807 Generator loss: 3.0621
Epoch 0/1... Batch 70/6331... Discriminator loss: 1.5137 Generator loss: 1.0864
Epoch 0/1... Batch 80/6331... Discriminator loss: 1.1109 Generator loss: 0.8963
Epoch 0/1... Batch 90/6331... Discriminator loss: 1.6350 Generator loss: 3.3098
Epoch 0/1... Batch 100/6331... Discriminator loss: 2.0212 Generator loss: 0.5976

Batches:   2%|█▏                                                                       | 100/6331 [00:14<14:46,  7.03batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 1.6211 Generator loss: 2.5513
Epoch 0/1... Batch 120/6331... Discriminator loss: 2.5464 Generator loss: 2.2082
Epoch 0/1... Batch 130/6331... Discriminator loss: 3.7240 Generator loss: 0.0723
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.4602 Generator loss: 0.5043
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.1416 Generator loss: 1.3309
Epoch 0/1... Batch 160/6331... Discriminator loss: 2.3332 Generator loss: 1.2819
Epoch 0/1... Batch 170/6331... Discriminator loss: 1.7067 Generator loss: 0.6473
Epoch 0/1... Batch 180/6331... Discriminator loss: 1.7837 Generator loss: 0.5620
Epoch 0/1... Batch 190/6331... Discriminator loss: 1.6740 Generator loss: 0.6191
Epoch 0/1... Batch 200/6331... Discriminator loss: 1.4184 Generator loss: 0.8758

Batches:   3%|██▎                                                                      | 200/6331 [00:27<14:07,  7.23batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.5115 Generator loss: 0.6955
Epoch 0/1... Batch 220/6331... Discriminator loss: 2.0329 Generator loss: 0.4020
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.3306 Generator loss: 0.8538
Epoch 0/1... Batch 240/6331... Discriminator loss: 1.4250 Generator loss: 0.9303
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.2684 Generator loss: 0.9046
Epoch 0/1... Batch 260/6331... Discriminator loss: 2.0126 Generator loss: 0.4926
Epoch 0/1... Batch 270/6331... Discriminator loss: 1.6497 Generator loss: 0.6497
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.6938 Generator loss: 0.7078
Epoch 0/1... Batch 290/6331... Discriminator loss: 1.6415 Generator loss: 0.6343
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.3235 Generator loss: 0.9469

Batches:   5%|███▍                                                                     | 300/6331 [00:40<13:36,  7.38batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.6571 Generator loss: 0.5934
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.6070 Generator loss: 0.5833
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.5757 Generator loss: 0.6724
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.5104 Generator loss: 0.9249
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.6844 Generator loss: 0.6001
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.4761 Generator loss: 0.6771
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.5973 Generator loss: 0.6548
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.3979 Generator loss: 0.8049
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.5500 Generator loss: 0.7721
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.4877 Generator loss: 0.6667

Batches:   6%|████▌                                                                    | 400/6331 [00:53<13:19,  7.42batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.5855 Generator loss: 0.5881
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.3911 Generator loss: 0.7343
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.5023 Generator loss: 0.5793
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.4787 Generator loss: 0.7153
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.4467 Generator loss: 0.7697
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.4534 Generator loss: 0.6112
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.3819 Generator loss: 0.8482
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.5866 Generator loss: 0.6314
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.7906 Generator loss: 0.5196
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.4754 Generator loss: 0.7509

batch_size = 32
z_dim = 200
learning_rate = 0.001
beta1 = 0.3 # the losses seem to vary less, but the visuals are still bad
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 10.4056 Generator loss: 0.0001
Epoch 0/1... Batch 20/6331... Discriminator loss: 4.0650 Generator loss: 0.0846
Epoch 0/1... Batch 30/6331... Discriminator loss: 2.5723 Generator loss: 0.1731
Epoch 0/1... Batch 40/6331... Discriminator loss: 1.8835 Generator loss: 0.3129
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.8609 Generator loss: 0.5904
Epoch 0/1... Batch 60/6331... Discriminator loss: 2.0750 Generator loss: 1.4860
Epoch 0/1... Batch 70/6331... Discriminator loss: 2.3003 Generator loss: 0.2562
Epoch 0/1... Batch 80/6331... Discriminator loss: 1.8901 Generator loss: 0.4367
Epoch 0/1... Batch 90/6331... Discriminator loss: 1.8324 Generator loss: 0.6193
Epoch 0/1... Batch 100/6331... Discriminator loss: 1.6307 Generator loss: 0.6006

Batches:   2%|█▏                                                                       | 100/6331 [00:13<14:11,  7.32batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 1.4872 Generator loss: 0.9883
Epoch 0/1... Batch 120/6331... Discriminator loss: 1.6589 Generator loss: 0.6321
Epoch 0/1... Batch 130/6331... Discriminator loss: 1.6368 Generator loss: 0.6155
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.6559 Generator loss: 0.5191
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.6236 Generator loss: 0.6155
Epoch 0/1... Batch 160/6331... Discriminator loss: 1.5269 Generator loss: 0.6294
Epoch 0/1... Batch 170/6331... Discriminator loss: 1.8081 Generator loss: 0.5218
Epoch 0/1... Batch 180/6331... Discriminator loss: 1.8372 Generator loss: 0.4246
Epoch 0/1... Batch 190/6331... Discriminator loss: 1.5381 Generator loss: 0.7126
Epoch 0/1... Batch 200/6331... Discriminator loss: 1.5079 Generator loss: 0.6517

Batches:   3%|██▎                                                                      | 200/6331 [00:27<13:59,  7.31batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.5709 Generator loss: 0.6491
Epoch 0/1... Batch 220/6331... Discriminator loss: 1.5871 Generator loss: 0.7032
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.4959 Generator loss: 0.7392
Epoch 0/1... Batch 240/6331... Discriminator loss: 1.5441 Generator loss: 0.5996
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.6637 Generator loss: 0.5064
Epoch 0/1... Batch 260/6331... Discriminator loss: 1.5341 Generator loss: 0.6727
Epoch 0/1... Batch 270/6331... Discriminator loss: 1.5182 Generator loss: 0.8540
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.4799 Generator loss: 0.7007
Epoch 0/1... Batch 290/6331... Discriminator loss: 1.4984 Generator loss: 0.8754
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.3619 Generator loss: 0.7930

Batches:   5%|███▍                                                                     | 300/6331 [00:41<13:48,  7.28batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.5283 Generator loss: 0.6061
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.4871 Generator loss: 0.6821
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.4720 Generator loss: 0.7444
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.4548 Generator loss: 0.7370
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.4864 Generator loss: 0.6868
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.5173 Generator loss: 0.6263
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.5234 Generator loss: 0.6188
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.5922 Generator loss: 0.5785
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.4578 Generator loss: 0.7792
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.5291 Generator loss: 0.7264

Batches:   6%|████▌                                                                    | 400/6331 [00:57<14:13,  6.95batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.5917 Generator loss: 0.6207
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.5245 Generator loss: 0.7219
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.5410 Generator loss: 0.5125
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.7262 Generator loss: 0.5531
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.4264 Generator loss: 0.7698
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.5525 Generator loss: 0.5575
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.5002 Generator loss: 0.6173
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.5452 Generator loss: 0.8027
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.5993 Generator loss: 0.6056
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.5972 Generator loss: 0.7396

batch_size = 32
z_dim = 200
learning_rate = 0.002
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 5.4785 Generator loss: 0.0145
Epoch 0/1... Batch 20/6331... Discriminator loss: 3.1037 Generator loss: 0.2022
Epoch 0/1... Batch 30/6331... Discriminator loss: 1.5078 Generator loss: 3.5222
Epoch 0/1... Batch 40/6331... Discriminator loss: 1.8756 Generator loss: 0.3056
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.4507 Generator loss: 0.6708
Epoch 0/1... Batch 60/6331... Discriminator loss: 2.0680 Generator loss: 3.3547
Epoch 0/1... Batch 70/6331... Discriminator loss: 1.1757 Generator loss: 0.8846
Epoch 0/1... Batch 80/6331... Discriminator loss: 1.7818 Generator loss: 0.3975
Epoch 0/1... Batch 90/6331... Discriminator loss: 1.9203 Generator loss: 0.4779
Epoch 0/1... Batch 100/6331... Discriminator loss: 1.3721 Generator loss: 1.3006

Batches:   2%|█▏                                                                       | 100/6331 [00:14<15:02,  6.90batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 2.1100 Generator loss: 2.6391
Epoch 0/1... Batch 120/6331... Discriminator loss: 1.0850 Generator loss: 1.1347
Epoch 0/1... Batch 130/6331... Discriminator loss: 1.2320 Generator loss: 1.1427
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.9154 Generator loss: 0.4311
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.6456 Generator loss: 0.6091
Epoch 0/1... Batch 160/6331... Discriminator loss: 1.3764 Generator loss: 0.5449
Epoch 0/1... Batch 170/6331... Discriminator loss: 1.9273 Generator loss: 1.0420
Epoch 0/1... Batch 180/6331... Discriminator loss: 2.5397 Generator loss: 0.1762
Epoch 0/1... Batch 190/6331... Discriminator loss: 1.8594 Generator loss: 0.7574
Epoch 0/1... Batch 200/6331... Discriminator loss: 1.3051 Generator loss: 0.6759

Batches:   3%|██▎                                                                      | 200/6331 [00:28<14:34,  7.01batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.8203 Generator loss: 0.5784
Epoch 0/1... Batch 220/6331... Discriminator loss: 2.0253 Generator loss: 1.1097
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.8859 Generator loss: 0.6231
Epoch 0/1... Batch 240/6331... Discriminator loss: 1.9620 Generator loss: 0.4044
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.4035 Generator loss: 0.9282
Epoch 0/1... Batch 260/6331... Discriminator loss: 1.1818 Generator loss: 0.9234
Epoch 0/1... Batch 270/6331... Discriminator loss: 2.4885 Generator loss: 0.4937
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.7995 Generator loss: 1.0746
Epoch 0/1... Batch 290/6331... Discriminator loss: 2.1149 Generator loss: 1.3782
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.3965 Generator loss: 0.7276

Batches:   5%|███▍                                                                     | 300/6331 [00:42<14:15,  7.05batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.6824 Generator loss: 0.4958
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.6539 Generator loss: 0.9383
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.6468 Generator loss: 0.5177
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.5757 Generator loss: 0.4954
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.6905 Generator loss: 0.5806
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.7350 Generator loss: 0.4691
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.2261 Generator loss: 0.8629
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.4403 Generator loss: 0.6781
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.4816 Generator loss: 0.6680
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.5251 Generator loss: 0.7345

Batches:   6%|████▌                                                                    | 400/6331 [00:55<13:47,  7.17batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.3398 Generator loss: 0.9800
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.2415 Generator loss: 0.7549
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.4504 Generator loss: 0.8302
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.3993 Generator loss: 0.7752
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.5410 Generator loss: 0.7830
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.5340 Generator loss: 0.6813
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.6678 Generator loss: 0.4724
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.5147 Generator loss: 0.5970
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.7731 Generator loss: 0.9161
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.6655 Generator loss: 0.5698

batch_size = 32
z_dim = 200
learning_rate = 0.002
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 3.0599 Generator loss: 0.2022
Epoch 0/1... Batch 20/6331... Discriminator loss: 1.6020 Generator loss: 1.0375
Epoch 0/1... Batch 30/6331... Discriminator loss: 2.0318 Generator loss: 0.3676
Epoch 0/1... Batch 40/6331... Discriminator loss: 2.2556 Generator loss: 3.6343
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.2799 Generator loss: 0.8415
Epoch 0/1... Batch 60/6331... Discriminator loss: 1.4532 Generator loss: 3.6584
Epoch 0/1... Batch 70/6331... Discriminator loss: 1.1448 Generator loss: 1.1149
Epoch 0/1... Batch 80/6331... Discriminator loss: 4.1130 Generator loss: 1.9821
Epoch 0/1... Batch 90/6331... Discriminator loss: 1.7820 Generator loss: 0.5212
Epoch 0/1... Batch 100/6331... Discriminator loss: 1.3914 Generator loss: 0.9032

Batches:   2%|█▏                                                                       | 100/6331 [00:20<21:13,  4.89batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 2.1552 Generator loss: 0.6223
Epoch 0/1... Batch 120/6331... Discriminator loss: 1.6587 Generator loss: 0.4743
Epoch 0/1... Batch 130/6331... Discriminator loss: 2.2416 Generator loss: 0.2690
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.8161 Generator loss: 0.4261
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.8663 Generator loss: 1.2373
Epoch 0/1... Batch 160/6331... Discriminator loss: 1.7989 Generator loss: 0.5121
Epoch 0/1... Batch 170/6331... Discriminator loss: 2.2521 Generator loss: 0.6004
Epoch 0/1... Batch 180/6331... Discriminator loss: 1.8970 Generator loss: 0.6565
Epoch 0/1... Batch 190/6331... Discriminator loss: 1.6649 Generator loss: 0.5746
Epoch 0/1... Batch 200/6331... Discriminator loss: 3.0268 Generator loss: 0.2713

Batches:   3%|██▎                                                                      | 200/6331 [00:40<20:38,  4.95batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.4654 Generator loss: 1.1791
Epoch 0/1... Batch 220/6331... Discriminator loss: 1.8661 Generator loss: 1.2836
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.5679 Generator loss: 0.5172
Epoch 0/1... Batch 240/6331... Discriminator loss: 1.6973 Generator loss: 0.6153
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.6238 Generator loss: 0.9990
Epoch 0/1... Batch 260/6331... Discriminator loss: 1.6379 Generator loss: 1.0291
Epoch 0/1... Batch 270/6331... Discriminator loss: 1.9137 Generator loss: 0.4852
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.7044 Generator loss: 0.7043
Epoch 0/1... Batch 290/6331... Discriminator loss: 1.6377 Generator loss: 0.7749
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.5941 Generator loss: 0.6610

Batches:   5%|███▍                                                                     | 300/6331 [01:01<20:37,  4.87batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.8440 Generator loss: 0.6588
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.7761 Generator loss: 0.5778
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.6315 Generator loss: 0.6715
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.8651 Generator loss: 1.1667
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.7157 Generator loss: 0.3781
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.4067 Generator loss: 0.8118
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.9382 Generator loss: 0.5852
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.3596 Generator loss: 0.8081
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.5816 Generator loss: 0.6382
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.5895 Generator loss: 0.7250

Batches:   6%|████▌                                                                    | 400/6331 [01:21<20:03,  4.93batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.5304 Generator loss: 0.7694
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.6198 Generator loss: 0.6161
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.5929 Generator loss: 0.9247
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.5276 Generator loss: 0.6868
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.6165 Generator loss: 0.6713
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.4023 Generator loss: 0.9631
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.5503 Generator loss: 0.5960
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.4513 Generator loss: 0.7274
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.4405 Generator loss: 0.7560
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.4869 Generator loss: 0.7936

batch_size = 32
z_dim = 200
learning_rate = 0.002
beta1 = 0.3
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

Epochs:   0%|                                                                                       | 0/1 [00:00<?, ?epoch/s]
Batches:   0%|                                                                                   | 0/6331 [00:00<?, ?batch/s]

Epoch 0/1... Batch 10/6331... Discriminator loss: 3.9754 Generator loss: 0.1650
Epoch 0/1... Batch 20/6331... Discriminator loss: 2.0891 Generator loss: 0.2722
Epoch 0/1... Batch 30/6331... Discriminator loss: 1.5292 Generator loss: 1.1602
Epoch 0/1... Batch 40/6331... Discriminator loss: 1.3910 Generator loss: 1.4078
Epoch 0/1... Batch 50/6331... Discriminator loss: 1.5541 Generator loss: 0.6097
Epoch 0/1... Batch 60/6331... Discriminator loss: 3.4607 Generator loss: 2.9490
Epoch 0/1... Batch 70/6331... Discriminator loss: 1.4156 Generator loss: 1.1056
Epoch 0/1... Batch 80/6331... Discriminator loss: 1.3131 Generator loss: 1.4627
Epoch 0/1... Batch 90/6331... Discriminator loss: 1.3870 Generator loss: 0.8783
Epoch 0/1... Batch 100/6331... Discriminator loss: 1.4284 Generator loss: 0.5692

Batches:   2%|█▏                                                                       | 100/6331 [00:13<14:18,  7.26batch/s]

Epoch 0/1... Batch 110/6331... Discriminator loss: 2.3961 Generator loss: 0.3325
Epoch 0/1... Batch 120/6331... Discriminator loss: 1.6608 Generator loss: 0.6194
Epoch 0/1... Batch 130/6331... Discriminator loss: 1.7591 Generator loss: 0.6958
Epoch 0/1... Batch 140/6331... Discriminator loss: 1.5826 Generator loss: 0.5292
Epoch 0/1... Batch 150/6331... Discriminator loss: 1.5203 Generator loss: 0.6302
Epoch 0/1... Batch 160/6331... Discriminator loss: 1.6977 Generator loss: 0.5293
Epoch 0/1... Batch 170/6331... Discriminator loss: 2.3793 Generator loss: 0.8690
Epoch 0/1... Batch 180/6331... Discriminator loss: 1.4595 Generator loss: 0.5302
Epoch 0/1... Batch 190/6331... Discriminator loss: 2.3990 Generator loss: 0.2592
Epoch 0/1... Batch 200/6331... Discriminator loss: 1.4603 Generator loss: 0.8557

Batches:   3%|██▎                                                                      | 200/6331 [00:27<13:56,  7.33batch/s]

Epoch 0/1... Batch 210/6331... Discriminator loss: 1.6008 Generator loss: 0.7085
Epoch 0/1... Batch 220/6331... Discriminator loss: 1.5260 Generator loss: 0.7477
Epoch 0/1... Batch 230/6331... Discriminator loss: 1.4714 Generator loss: 0.8830
Epoch 0/1... Batch 240/6331... Discriminator loss: 1.5238 Generator loss: 0.6388
Epoch 0/1... Batch 250/6331... Discriminator loss: 1.5442 Generator loss: 0.6673
Epoch 0/1... Batch 260/6331... Discriminator loss: 1.7661 Generator loss: 0.5013
Epoch 0/1... Batch 270/6331... Discriminator loss: 1.4713 Generator loss: 0.9840
Epoch 0/1... Batch 280/6331... Discriminator loss: 1.4686 Generator loss: 0.6756
Epoch 0/1... Batch 290/6331... Discriminator loss: 1.6939 Generator loss: 0.6037
Epoch 0/1... Batch 300/6331... Discriminator loss: 1.3958 Generator loss: 0.8836

Batches:   5%|███▍                                                                     | 300/6331 [00:41<14:01,  7.16batch/s]

Epoch 0/1... Batch 310/6331... Discriminator loss: 1.5609 Generator loss: 0.5396
Epoch 0/1... Batch 320/6331... Discriminator loss: 1.4529 Generator loss: 0.6442
Epoch 0/1... Batch 330/6331... Discriminator loss: 1.6195 Generator loss: 0.7108
Epoch 0/1... Batch 340/6331... Discriminator loss: 1.4886 Generator loss: 0.7618
Epoch 0/1... Batch 350/6331... Discriminator loss: 1.5131 Generator loss: 0.6900
Epoch 0/1... Batch 360/6331... Discriminator loss: 1.4492 Generator loss: 0.6242
Epoch 0/1... Batch 370/6331... Discriminator loss: 1.4766 Generator loss: 0.7120
Epoch 0/1... Batch 380/6331... Discriminator loss: 1.5052 Generator loss: 0.7291
Epoch 0/1... Batch 390/6331... Discriminator loss: 1.4652 Generator loss: 0.7663
Epoch 0/1... Batch 400/6331... Discriminator loss: 1.4861 Generator loss: 0.7472

Batches:   6%|████▌                                                                    | 400/6331 [00:57<14:17,  6.92batch/s]

Epoch 0/1... Batch 410/6331... Discriminator loss: 1.5733 Generator loss: 0.6276
Epoch 0/1... Batch 420/6331... Discriminator loss: 1.6487 Generator loss: 0.5782
Epoch 0/1... Batch 430/6331... Discriminator loss: 1.4759 Generator loss: 0.6065
Epoch 0/1... Batch 440/6331... Discriminator loss: 1.4839 Generator loss: 0.6674
Epoch 0/1... Batch 450/6331... Discriminator loss: 1.5895 Generator loss: 0.7082
Epoch 0/1... Batch 460/6331... Discriminator loss: 1.5227 Generator loss: 0.7739
Epoch 0/1... Batch 470/6331... Discriminator loss: 1.6212 Generator loss: 0.4643
Epoch 0/1... Batch 480/6331... Discriminator loss: 1.4399 Generator loss: 1.0099
Epoch 0/1... Batch 490/6331... Discriminator loss: 1.5532 Generator loss: 0.6102
Epoch 0/1... Batch 500/6331... Discriminator loss: 1.4458 Generator loss: 0.7662

batch_size = 32
z_dim = 200
learning_rate = 0.01 # too high
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

batch_size = 32
z_dim = 200
learning_rate = 0.004 # too high
beta1 = 0.5
alpha = 0.2

"""
DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE
"""
epochs = 1

celeba_dataset = helper.Dataset('celeba', glob(os.path.join(data_dir, 'img_align_celeba/*.jpg')))
with tf.Graph().as_default():
    train(epochs, batch_size, z_dim, learning_rate, beta1, celeba_dataset.get_batches,
          celeba_dataset.shape, celeba_dataset.image_mode, alpha)

z_dim = 200
beta1 = 0.5

# Generate some random faces
def generate_faces(checkpoint=None, z_dim=z_dim, beta1=beta1, show_n_images=25):
    graph = tf.Graph()
    if not checkpoint:
        checkpoint = tf.train.latest_checkpoint('checkpoints')
    print('Generating faces from saved checkpoint:', checkpoint)
    with graph.as_default():
        input_real, input_z, learn_rate = model_inputs(28, 28, 3, z_dim)
        d_loss, g_loss = model_loss(input_real, input_z, 3)
        d_opt, g_opt = model_opt(d_loss, g_loss, learn_rate, beta1)
        #dummy = tf.Variable(0)
        saver = tf.train.Saver()
        with tf.Session() as sess:
            #sess.run(tf.global_variables_initializer())
            saver.restore(sess, checkpoint)
            show_generator_output(sess, show_n_images, input_z, 3, 'RGB')

generate_faces()

Generating faces from saved checkpoint: checkpoints\generator_bs32_zd200_lr0.002_b0.3.ckpt

generate_faces(show_n_images=16)

Generating faces from saved checkpoint: checkpoints\generator_bs32_zd200_lr0.002_b0.3.ckpt

So far the hyperparameters that give better looking visuals

batch_size = 32
z_dim = 200
learning_rate = 0.002
beta1 = 0.5
alpha = 0.2

generate_faces('checkpoints\generator_bs32_zd200_lr0.002_b0.5.ckpt')

Generating faces from saved checkpoint: checkpoints\generator_bs32_zd200_lr0.002_b0.5.ckpt

generate_faces('checkpoints\generator_bs32_zd200_lr0.002_b0.5.ckpt', show_n_images=16)

Generating faces from saved checkpoint: checkpoints\generator_bs32_zd200_lr0.002_b0.5.ckpt

Submitting This Project¶

When submitting this project, make sure to run all the cells before saving the notebook. Save the notebook file as "dlnd_face_generation.ipynb" and save it as a HTML file under "File" -> "Download as". Include the "helper.py" and "problem_unittests.py" files in your submission.